GAN artifacts in image corners

Hello, pytorch community!

I’m currently training GAN models and totally satisfied with it’s result, but it has a very annoying bug. In corners of the image (mostly in bottom right, but others appear sometimes as well) I see strange pattern that model generates. Those artifacts take a small portion of 1024x1024 (maybe 1/16th) image but obviously ruin the perception.

Снимок экрана 2020-05-24 в 16.49.20 image image

I’m using Pix2PixHD-like network with MultiscaleDiscriminator.

How one should debug it? I think zero-padding may be involved here, but I am afraid that reflect padding will lose spatial info. Are there any recommendations on how to fight this? I think that pointing discriminator’s attention to corners somehow may work, but I have no idea how to approach that